Daily Briefing
18 November 2026 · 5 stories
🔥 Top story
1Tencent opens the Hy4 flash weights: the licence lets us work
The weights of Tencent's Hy4 flash model are finally available: the repository holds 148 weight files, and you can download them. The interesting part is not technical, it is the contract. The licence read inside the repository is not the usual free one: it adds two thresholds. Above 50 million dollars of worldwide revenue, or 100 million monthly active users, you need a written authorisation from the publisher. Below that, use is free, and Norway appears in no exclusion clause, unlike several recent Chinese models. The model counts 220 billion parameters in total, of which 11.3 billion work on every word. In plain terms, it is a rental with two price tiers: stay under the threshold and you pay the normal rate. Source : https://huggingface.co/tencent/Hy4-flash
2OpenAI cuts the price of text already read, and teamwork becomes native
OpenAI updated its two models, Astra for hard tasks and Sol for everyday work. The number that changes the bill on long sessions: text that has already been read now costs 0.40 dollars instead of 1, so 60 % less. That is exactly what you pay when an agent keeps re-reading the same file. The other change is quieter: the official interface can now hand parts of the job to helper agents, with no homemade workaround. The most powerful model drops to 8 dollars per million words read, against 10 before. For a service that sells a finished result rather than tokens, that is the cost line that matters. Source : https://platform.openai.com/docs/models/gpt-6.2-astra
3Anthropic lets your application bill the end customer
For the first time, an application can charge the customer for their own Claude usage. The customer links their account, the application picks at each call who pays, and the customer sets a monthly cap. If the debit fails, the request is refused rather than billed to the developer. Two limits to know: only customers who pay Anthropic directly benefit, and the customer's confidentiality guarantees no longer apply to those requests. The picture is simple: it is the till receipt of the access provider, but it belongs to the customer. Source : https://docs.anthropic.com/en/release-notes/api
4A study puts a price on splitting work: fewer conflicts caught
When you split research across several agents, they bring back more facts that contradict each other: 50.2 % against 68.7 %. But they catch fewer: 16.2 % against 13.2 %. The result looks paradoxical and is not: detection is better when a single agent saw both statements, and worse when they are split. In other words, the number of agents is not the problem, the separation of information is. The authors conclude that handing information from the worker to the leader is the priority target. For us, the rule is simple: have the versions confronted side by side by whoever saw both. Source : https://arxiv.org/abs/2611.22245
5Model ML raises 250 million dollars with agents at 2 % of the price
The London company Model ML raises 250 million dollars, valuing the business at 1.5 billion. It automates the repetitive work of investment bankers: financial models, slide decks, reports. Its economic argument is a number: its agents run on open models and operate at about 2 % of the cost of frontier models, according to the publisher, with no independent measurement. The rest is a question of business model: the agent does not sell tokens, it sells finished work. That is the difference between charging for consumption and charging for a result. Source : https://techcrunch.com/2026/11/17/model-ml-raises-250m-to-automate-investment-banking-grunt-work/
📡 To watch
Microsoft prepares the company context for every agent
Microsoft introduces IQ, a layer that gathers the work of people, company data, reference documents and the web, with more than 350 connectors from third parties. The interesting point: a company module prepared inside Microsoft can be reused by other assistants, including those of OpenAI and Anthropic. Prepare the context once, use it everywhere. Source : https://azure.microsoft.com/en-us/blog/introducing-the-new-microsoft-iq-putting-your-organizations-intelligence-to-work-in-every-agent/
NVIDIA stops selling chips to China and watches its exports
Two announcements on the same day. Management said on the earnings call that the company no longer ships data centre chips to China, even though Washington had allowed it: 2.4 billion dollars were sold in the previous quarter, and none are expected in the next. At the same time, the board creates a permanent committee for national security and export controls, meeting at least once a quarter. Source : https://nvidianews.nvidia.com/news/nvidia-board-of-directors-establishes-national-security-and-export-control-oversight-committee
Liquid AI measures the confidence error to 0.8 of a point
These models do not write text: they return a probability for every possible answer, and they now accept audio. The calibration error reaches 0.8 points on thirteen test suites, against 3.7 for the Perplexity equivalent and 7.2 for a frontier model. When the model says 80 %, it is right 80 % of the time, so a threshold for handing over to a human can be set. Source : https://www.liquid.ai/blog/d1-2
Datadog ties production tests to the agent bill
You can split traffic at random between two versions of an agent and measure what really changes: the cost, the speed, the errors. The published example: switching models cut the estimated cost by 44.3 % without any loss of quality, and the tool calls that failed were used to fix the agent before deployment. Source : https://www.infoq.com/news/2026/11/datadog-ai-agent-observability/
📊 Trend
Today's news tells one story: agents are leaving the lab and moving into organisations, with consequences you can measure. A Chinese publisher releases the weights of its model with a licence that lets small companies use it. Another one cuts the price of text that has already been read, because that is where the bill of a long session hides. A third lets the application charge the end customer, which changes how an agent can be sold. And a study puts a number on what we suspected without measuring: splitting work across agents brings back more contradictions and catches fewer. In practice, your agents cost less to run, but they demand vigilance in how information is shared between them.